-) properly depending on libraries.
-) adding an independent 'loader' class to use instead of require 'grpc/grpc'.

improve code reuse

for now. This will increase our coverage of concurrent execution cases in C++
since we don't currently have any coverage for multithreaded async execution using
a common completion queue.

In this test, 100 client threads initiate 1000 RPCs each, then join, and then
100 client threads come along to proces those RPC completion events as they come back.

Importantly:
- allows matching timeout values between makefile and run_tests without
  repeating ourselves
- allows borrowing of compiler flags by other build systems

There's still a little too much build configuration built into our
Makefile, but we can start attacking that over time.